PARAS MADAN
All writing

Latest in AI

Sidekick Builder: A $7-a-Week Robotics Development Workspace

Sidekick's self-serve beta combines model discovery, routing and training, with important limits around inference costs and robot integration.

Sidekick Robotics introduced the beta of Builder on October 23, giving engineers a self-serve workspace for its physical AI platform. The plan brings model discovery, routing and training into one environment, with a first week free and a $7 weekly subscription afterward.

For a small robotics team, the appeal is having somewhere to compare models before committing engineering time to an integration. A model's output format, hardware requirements and availability can matter as much as its performance on a demonstration task.

What the weekly subscription covers

The pricing page lists Builder as available, alongside a free, view-only Sandbox. Builder includes model use, routing, training and fine-tuning. Team is marked coming soon, while Enterprise has custom pricing and options such as on-premise infrastructure and dedicated support.

The $7 price should be read as the workspace subscription. Sidekick separately describes inference billing by action step, predicted frame or token, depending on the operation. A developer comparing costs needs to account for the model calls a workload will generate, including repeated attempts during development.

A short experiment would therefore need two budgets: the recurring workspace charge and the inference spend. The public pricing summary does not specify an included inference allowance, so I would check that before estimating a project's total cost.

Choosing a model before running it

Builder's Model Hub exposes information including licenses, memory requirements and whether a model has been tested on a real robot. It distinguishes live offerings from open weights, closed models and research entries. Developers can shortlist up to 12 models per project and preview a route before execution.

That distinction is useful because appearing in a catalog does not establish that a model is callable on the hardware in front of you. The selection process still needs to answer where the model runs and whether it supports the robot's observations and action space.

The API documentation describes exact checkpoint selection as well as aliases that let the router choose a model. It also documents per-call cost ceilings, provider restrictions and route previews that estimate costs without executing inference. Responses record attempted providers and fallbacks, so a developer can inspect which route produced an answer.

For an experiment comparing policies, I would start with a fixed checkpoint and retain those traces. Otherwise, a change in the selected model could be mistaken for an improvement in the application code.

The robot integration still needs attention

Sidekick's earlier SDK engineering guide explains a constraint that remains relevant to Builder: camera names must match the selected checkpoint, and joint-state inputs must have the expected width and ordering. A correctly formatted request can still describe the scene incorrectly if those inputs are mismatched.

The guide also describes an action-streaming client that fetches future action chunks in the background. When the buffer is empty or a chunk is stale, the controller should hold or decelerate. This is an integration detail worth testing with the hardware, especially when a remote model or network connection becomes slow.

Builder is also a narrower purchase than the complete platform. The launch announcement places Learning Loops, Flywheel, Private Model and Evals behind an Enterprise upgrade. Teams planning around those features need to resolve that commercial scope first.

I would judge the beta on a contained workflow: select one supported robot and task, verify the observation contract, inspect a route preview and measure the cost of repeated trials. That would provide a more useful basis for adoption than the number of models listed in the Hub.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn robotics infrastructure, product launches and integration details into clear stories that developers can understand and share.

Book a 30-minute call.